Training method and training device for training ai system for classifying medical images

Abstract

The invention relates to a training method for training an AI system for classifying medical images, the AI system comprising a machine learning pre-classifier model and a machine learning classifier unit, wherein several feature vectors of layers preceding an output layer of the pre-classifier model are input to the classifier unit, the training method comprising the following steps: - providing training data, the training data comprising a plurality of 2D and / or 3D images and ground truths, the plurality of 2D and / or 3D images showing organs in a human body, the ground truths comprising a classification of the organs in the images, - training the pre-classifier model with the training data to provide a classification of predefined organs shown in the images in accordance with the ground truths of the training data, - training the classifier unit by inputting the training data into the trained pre-classifier model so as to provide a classification of predefined organs shown in the images by the classifier unit in accordance with the ground truths of the training data. Furthermore, the invention describes a training device, an AI system and a medical system.

Description

Technical Field

[0001] This invention describes a training method and training device for training an AI system for classifying medical images, an AI system for classifying medical images, and a medical system. Background Technology

[0002] Organ and pathology identification is crucial for automating various processes in medical imaging workflows. One example is utilizing appropriate lesion segmentation algorithms for user-provided organ localization. Organ segmentation models naturally serve as a baseline for these types of applications. One limitation of organ segmentation is computation time, which can be too high for interactive user experiences. Additionally, for some applications, the architectural overhead for preprocessing may outweigh the benefits, rendering them impractical.

[0003] Medical image segmentation has undergone continuous improvements in the literature (see, for example, "Unetr: Transformers for 3d medical imagesegmentation" by Hatamizadeh, Ali, et al., in the proceedings of the 2022 IEEE / CVF Winter Conference on Applications of Computer Vision). While accuracy has increased, so has the computational load. This is particularly true for 3D volumetric and transformer architectures, where processing images without a GPU becomes prohibitive. However, many applications only require a few points of interest, thus eliminating the need for complete segmentation.

[0004] Training a specific classifier for each use case has limitations in terms of aggregating training data and also incurs large runtime memory requirements. Conversely, adapting existing classifiers to new tasks is more appealing. This approach is referred to in the literature as base model and fine-tuning. However, until now, the problems discussed above have not been overcome. Summary of the Invention

[0005] The object of this invention is to improve upon known systems and methods, and to provide a training method and apparatus for training an AI system for classifying medical images, an AI system for classifying medical images, and a medical system, thereby overcoming the problems described above. A particular objective of this invention is to optimize the quality of image segmentation without requiring excessive computational resources.

[0006] This objective is achieved by the training method according to claim 1, the training device according to claim 9, the AI ​​system according to claim 10, and the medical system according to claim 13.

[0007] The training method according to the present invention is used to train an AI system for classifying medical images (specifically, medical image localization). The AI ​​system includes a machine learning pre-classifier model and machine learning classifier units. Several feature vectors of the layers preceding the output layer of the pre-classifier model are the inputs to the classifier units. The training method includes the following steps: - Provide training data, which includes multiple 2D and / or 3D images and ground truth values, wherein the multiple 2D and / or 3D images illustrate organs in the human body, and the ground truth values ​​include classifications of the organs in the images. - Train a pre-classifier model using the training data to provide classifications of predefined organs shown in the image based on the baseline ground truth of the training data. - The classifier unit is trained by feeding training data into a trained pre-classifier model so that the classifier unit can provide a classification of the predefined organs shown in the image based on the benchmark ground truth of the training data.

[0008] An important aspect of this invention is the special connection between the pre-classifier model and the classifier unit, as well as the separate training of the classifier unit.

[0009] A pre-classifier model is a machine learning model designed to segment an image by classifying its content. Such models are well known in the art. Preferably, the base classifier model used is also well known in the art. Base classifier models typically provide organ labels for predefined (particularly all) organs of the body. However, these labels are often general and may not be optimal for specific use cases (such as classifying organs like the lungs or liver) or other use cases for identifying lesions. Due to the independent nature of classifier models, variance near the boundary is high. Therefore, a pre-classifier model can classify organs, but not optimally. In summary, it can be said that a pre-classifier model is an organ classifier, not task-specific.

[0010] The classification of a pre-classifier model can now be optimized by taking its output (the classification vector) and feeding that output into another model, which further classifies or validates the output. The present invention offers an alternative approach by using the output of a layer preceding the pre-classifier model's output layer (i.e., several feature vectors) instead of the output of the final layer of the pre-classifier model (the classification vector).

[0011] Therefore, in addition to the pre-classifier model, the AI ​​system also includes a machine learning classifier unit, and at least one feature vector of the layer preceding the output layer of the pre-classifier model is the input of that classifier unit. The layer preceding the output layer can be any layer except the last layer; however, the penultimate layer is preferred because it contains the most information about the classification.

[0012] First, the pre-classifier model is trained, and then the classifier unit is trained. It should be noted that the pre-classifier model can be trained without completing the architecture of the AI ​​system. Therefore, a basic classification model can be taken, trained, and then the classifier unit added. While the final layer of the pre-classifier model (the output layer) is needed for training, it is no longer required in the AI ​​system and can be removed. However, it may be advantageous to use this output layer for additional classification, which can be used to validate or test the output of the classifier unit.

[0013] For training, training data is provided, comprising multiple 2D and / or 3D images and a ground truth value. The multiple 2D and / or 3D images illustrate organs in the human body, and the ground truth value includes the classification of the organs in the images. Each image may have its own ground truth value; however, a group of images with a common ground truth value may also be provided. This ground truth value may exist in the form of labels or in a segmented form. The images are preferably X-ray images, CT images, MRT images, and / or ultrasound images.

[0014] In a preferred embodiment, where the pre-classifier model is designed to classify images pixel by pixel and provide segmented images, the baseline ground truth may include a label for each pixel, including at least a label for each pixel showing an organ from a predefined list of organs.

[0015] Preferably, a first training dataset is used to train the pre-classifier model, and a second training dataset is used to train the classifier unit. The second dataset can be task-specific, for example, by using a baseline ground truth dataset that also relates to pathology such as lesions.

[0016] A pre-classifier model is trained using training data (e.g., using a first set) to provide a predefined (preferably all) classification of organs shown in the images of the training data, based on the ground truth values ​​of the training data. This training procedure is well-known and is often referred to as "supervised training." This training forces the feature vectors generated by the layers of the pre-classifier model to capture organ information. Any layer will capture a portion, but the last layer before the output layer is the closest. As described above, this training step is performed before training the classifier units. Therefore, the shape of the feature vectors and their informational content are well defined (and preferably not systematically changed because the weights of the pre-classifier model are fixed).

[0017] The basic model, which serves as a pre-classifier, is typically used for feature extraction. Feature extraction is based on the basic model's neural network. Particularly sparse sampling descriptors are used to prepare any point in the image for the basic model. This sparse sampling input improves speed due to the limited number of samples in the image space.

[0018] The architecture of the AI ​​system is established at least before training the classifier units. Then, any input data (image) passes through the trained pre-classifier model, but instead of the output label (the classification vector of the pre-classifier model), the hidden space representation prior to the label classifier is used as the feature vector(s). Preferably, any point in the input image is represented as an individual vector.

[0019] During the training of the classifier unit (“fine-tuning step”), training data, preferably the second set mentioned above, is input into the trained pre-classifier model so that the classifier unit provides a classification of predefined (preferably all) organs shown in the image based on the ground truth of the training data. This means that the pre-classifier model generates feature vectors of the input image, and at least one of these feature vectors is input into the classifier unit.

[0020] Here, specific training data can be generated for points provided in the supervision of the target task. In this (second) dataset, task-specific localizations and labels can be labeled as ground truth. Some random samples can also be added to the training to improve the robustness of the background class. After generating the training dataset, the classifier unit is trained to map basic model features to output labels.

[0021] Typically, the classifier units of the second independent neural network serving as the pre-classifier model are smaller than the pre-classifier model, allowing for faster training and requiring less data. Fine-tuning steps should not update the base model, ensuring it remains general. Due to the smaller size of the second classifier, multiple such classifiers can preferably be easily added together. Multiple classifier units can be used for different tasks and / or for the same task to obtain redundant results.

[0022] A preferred embodiment of the pre-classifier generating feature vectors for each pixel of an image is described, with a preferred and highly robust training method preferred. This is because a single feature vector is not as powerful as multiple feature vectors from the pre-classifier model. Therefore, features can be extended by adding neighboring localization features. For example, a 4 mm offset in both the x and y directions can be used, resulting in five features, including the center query point (like the number 5 on a die). The final representation of the localization (classification) can then be transformed into a concatenation of these features. Thus, the robustness of the classification is preferably improved by merging multiple nearby samples and using the most common labels, a simple rule to implement. This approach can be referred to as a "multi-call classifier." Here, the classifier unit is designed to take into account multiple feature vectors (of neighboring or nearby pixels) and process them to classify a single pixel (specifically the center pixel). The classification result is known from the baseline ground truth.

[0023] The resulting AI system can use sparse sampling descriptors and residual network classifiers (especially ResNet-1D). Training data can be generated from the Total Segmentator model, allowing the generated model to be used for distillation. As a result, the learned model becomes a general organ classifier.

[0024] Therefore, this invention relates to specific fine-tuning strategies. Instead of fine-tuning the entire AI system, two models are connected in a special architecture to adapt features extracted from a pre-classifier model (e.g., a base model) to a target task with enhanced features (particularly based on neighbors). Even for different tasks, the pre-classifier model can be the same, and only different (small) classifier units can be used, which are specifically trained for their individual tasks.

[0025] Beyond the algorithm selections mentioned here, automatically identifying organ context can be useful for other automation purposes. For example, it can be used to select CT Hounsfield windows, render preset selections, and filter existing findings based on given locations. Many more applications can be added with minimal overhead.

[0026] The training apparatus according to the present invention is used to train an AI system for classifying medical images. The AI ​​system includes a machine learning pre-classifier model and a machine learning classifier unit, wherein the feature vector of the layer preceding the output layer of the pre-classifier model is the input to the classifier unit. The training apparatus includes the following components: - A data interface designed to receive training data comprising multiple 2D and / or 3D images and ground truth values, wherein the multiple 2D and / or 3D images illustrate organs in the human body, and the ground truth values ​​include classifications of the organs in the images. - Pre-trained units, designed to train a pre-classifier model using training data to provide classifications of predefined organs as shown in the image. - A classifier training unit, designed to train a classifier unit using training data input into a trained pre-classifier model, to provide classification of predefined organs as shown in the image. The training equipment is designed to train the pre-classifier model and classifier unit using the training method according to the present invention.

[0027] The functions of the components of the training device are as described above.

[0028] Data interfaces are well-known in this field. They must be able to receive training data. They can be interfaces to data networks or databases.

[0029] The pre-training unit and the classifier training unit can be separate hardware devices (computers) or modules within a computing unit. Preferably, they are two different devices because the architecture and training behavior of the pre-classifier model and the classifier unit may differ. General units used for training machine learning models are well known in the art.

[0030] The AI ​​system according to the invention is used for classifying medical images. It includes a machine learning pre-classifier model and a machine learning classifier unit, wherein the layer preceding the output layer of the pre-classifier model is the input to the classifier unit, and the AI ​​system has been trained using the training method according to the invention. Specifically, the loss function is reduced based on a benchmark ground truth by optimizing the neural network parameters.

[0031] The architecture of the AI ​​system has been described above.

[0032] The medical systems according to the invention are, in particular, medical imaging systems, such as CT systems or MRT systems, or diagnostic systems, and include AI systems according to the invention.

[0033] This method can also be specified as a computer-implemented method. Some units or modules of the invention mentioned above can be implemented, wholly or partially, as software modules running on a processor of a computing system. Implementation primarily in the form of software modules has the advantage that applications already installed on existing computing systems can be updated with relatively little effort to install and run these units of the application. The object of the invention is also achieved by a computer program product having a computer program that can be directly loaded into the memory of a computing system, and the computer program including program units that, when executed by the computing system, perform the steps of the method, at least those steps that can be executed by a computer. In addition to the computer program, such a computer program product can also include other parts, such as documentation and / or additional components, and hardware components, such as hardware keys (dongles, etc.) to facilitate access to the software.

[0034] Computer-readable media (such as memory sticks, hard disks, or other transportable or permanently mounted carriers) can be used to transport and / or store executable portions of a computer program product, enabling these executable portions to be read from the processor unit of a computing system. The processor unit can include one or more microprocessors or equivalents thereof.

[0035] As shown in the following description, particularly advantageous embodiments and features of the invention are given by the dependent claims. Features from different claim classes may be appropriately combined to give additional embodiments not described herein.

[0036] According to a preferred training method, two different training datasets are used, wherein the pre-classifier model is trained using a first training dataset, and the classifier unit is trained using a second training dataset. As mentioned above, the first training dataset may include images of organs and general ground truth values ​​regarding the organs shown in the images. The second training dataset may be task-specific because it includes images illustrating specific scenarios and includes ground truth values ​​describing those specific scenarios. For example, it may be images illustrating pathologies with ground truth values ​​relating to those pathologies.

[0037] Preferably, the pre-classifier model is not trained while the classifier unit is being trained. Therefore, the weights of the pre-classifier model remain the same, and the feature vectors always have the same quality.

[0038] According to the preferred training method, the images of the training data are prepared using sparse sampling descriptors in order to train the pre-classifier model. Due to the limited number of samples in the image space, sparse sampling input improves speed. Sparse sampling descriptors are generally well known in the art (see, for example, "A Hierarchical Descriptor Framework for On-the-Fly Anatomical Location Matching between Longitudinal Studies," arXiv:2308.07337v2 [eess.IV], September 23, 2023, by Yerebakan H et al., or "Real Time Multi Organ Classification on Computed Tomography Images," arXiv:2404.18731v3 [cs.CV], January 9, 2025).

[0039] According to a preferred training method, the pre-classifier model is designed and trained to generate a feature vector for each pixel of the input image, and the classifier unit is designed to process multiple feature vectors. For each pixel of the image in the training data, the classifier unit is trained by taking the pixel as input and the feature vectors of multiple other pixels (particularly neighboring or nearby pixels) at predefined relative positions to the pixel, so that the classifier unit can provide a classification based on a ground truth value of the training data. The feature vectors can be concatenated to form information about the current query point and surrounding pixels. Preferably, a ResNet-1D architecture is used, which flattens the features and traverses the remaining connections. The use of multiple pixels enhances the robustness of the classifier unit.

[0040] According to a preferred training method, the pre-classifier model and / or classifier unit are trained by selecting a predefined region mask, moving the region mask over the training image, and selecting pixels at predefined locations within the region mask along with their ground truth values ​​for processing during training. For example, an offset of 4 mm is used in both the x and y directions, and five pixels are selected, one from the center of the offset and four from the corners of the offset. This produces five feature vectors, including features of the center query point. The final representation of the localization becomes a concatenation of these features.

[0041] According to the preferred training method, the training data, at least the training data used to train the classifier units, includes random samples. This is advantageous to improve the robustness of the background class. The sampled random locations are well-known and readily available. The pre-classifier model (e.g., the base model) is trained using those descriptors mapped to organ labels.

[0042] According to the preferred training method, the training data, particularly the second training data set used to train the classifier unit, includes images with predefined task-specific localizations and labeled images. This might be, for example, lung and liver lesions. In such cases, organ labels for given pathological localizations can be classified.

[0043] According to a preferred training method, the AI ​​system includes multiple machine learning classifier units, where the layer preceding the output layer of the pre-classifier model is the input to each classifier unit. Utilizing this advantageous "multi-call" principle, the same pre-classifier model (e.g., the same base classifier model) can be used in multiple scenarios. The pre-classifier model can be trained using only the output probability of an organ label. However, it is preferred that each classifier unit is trained using an individual training dataset. Therefore, each classifier unit can be optimized for a specific, individual task. It is also preferred that two or more classifier units are trained using the same or similar training data. This has the advantage of providing redundant results from classifiers, resulting in greater robustness. These classifier units can include different architectures. Such trained classifier units can be used to process the same input, and the outputs of these classifier units are processed statistically, for example, where the classification used has already been provided by most of these classifier units.

[0044] According to the preferred AI system, the last or penultimate feature layer before the output layer of the pre-classifier model is the input to the classifier unit. The input passes through the pre-classifier model, but instead of producing output labels, the hidden space representation before the label classifier is used here as the feature vector. It should be noted that each machine learning model has a well-defined output layer that outputs "understandable" features. These features can be given as ground truth for training. Layers preceding this output layer typically produce feature vectors that are not "understandable" (at least not in the case of the classification task involving this model). Therefore, even when the classification layer is missing, it is easy to determine whether a layer is the classification layer or the preceding layer.

[0045] According to the preferred AI system, the classifier unit is smaller than the pre-classifier model. This means that in this context, the classifier unit has smaller weights than the pre-classifier model. The dataset containing the weights may be less than 1 MB. Preferably, the AI ​​system includes multiple machine learning classifier units, where the layer preceding the output layer of the pre-classifier model is the input to each classifier unit. This allows for faster training and requires less data. Due to the smaller size of the second classifier, multiple such classifiers can be easily added together.

[0046] Generally speaking, artificial intelligence (machine learning models) is based on the principles of machine-based learning and is typically performed using adaptive algorithms that have been trained accordingly. The expression "machine learning" is often used for machine-based learning, which also includes the principles of "deep learning."

[0047] These methods may also include elements of "cloud computing." In the field of "cloud computing," IT infrastructure is provided through data networks, storage space, or processing power and / or application software. Communication between the user and the "cloud" is achieved via data interfaces and / or data transmission protocols. In the context of "cloud computing," in a preferred embodiment of the method according to the invention, data is provided to the "cloud" via a data channel (e.g., a data network). The "cloud" includes (remote) computing systems, such as computer clusters that typically do not include the user's local machine. Particularly preferred is that the cloud service provides computing power comparable to that of the application software. However, it should be noted that data transfer in the cloud may slow down execution. Attached Figure Description

[0048] Other objects and features of the invention will become apparent from the following detailed description taken in conjunction with the accompanying drawings. However, it should be understood that the drawings are for illustrative purposes only and are not intended to define limitations of the invention.

[0049] Figure 1 A training device according to the present invention is shown, which trains an AI system according to the present invention.

[0050] Figure 2 An AI system for classifying medical images according to the present invention is shown.

[0051] Figure 3 An AI system with two classifier units according to the present invention is shown.

[0052] Figure 4 A block diagram of the training method is shown.

[0053] Figure 5 A working example of the present invention is shown. Detailed Implementation

[0054] Figure 1 A training device 1 according to the invention is shown, which trains an AI system 5 according to the invention. The AI ​​system 5 includes a machine learning pre-classifier model 6 and a machine learning classifier unit 7, wherein multiple feature vectors of the layer preceding the output layer 8 of the pre-classifier model 6 are the inputs of the classifier unit 7. The output layer 8 is shown in dashed lines because it is required for training, not for the normal operation of the AI ​​system 5.

[0055] Training device 1 includes a data interface 2, a pre-training unit 3, and a classifier training unit. Data interface 2 is designed to receive training data T, which includes multiple 2D and / or 3D images D and a ground truth G. The multiple 2D and / or 3D images D illustrate organs in the human body, and the ground truth G includes the classification of the organs in image D (see also...). Figure 4 Pre-training unit 3 is designed to train pre-classifier model 6 using training data T to provide a classification C of the predefined organ shown in image D, and classifier training unit 4 is designed to train classifier unit 7 using training data T input into the trained pre-classifier model 6 to provide a classification C of the predefined organ shown in image D. In this example, two different training data sets T1 and T2 are used for training. Both exist in the training data T shown on the left. The second training data set T2 used to train classifier unit 7 may include predefined task-specific localization and labeled images D.

[0056] The symbolic network inside the box should only indicate that both pre-classifier model 6 and classifier unit 7 are machine learning models. The network of pre-classifier model 6 is symbolized as being larger than the network of classifier unit 7 to indicate that classifier unit 7 is smaller than pre-classifier model 6. Although the network is drawn as a fully connected network, other architectures may be used in practice. For example, the ResNet architecture, especially ResNet-1D. However, in any case, it is preferred that classifier unit 7 is smaller than pre-classifier model 6. This generally means fewer weights.

[0057] Figure 2 An AI system 5 for classifying a medical image D according to the present invention is shown. The AI ​​system 5 includes a machine learning pre-classifier model 6 and a machine learning classifier unit 7. The last layer before the output layer 8 (shown here in dashed lines) of the pre-classifier model 6 is the input to the classifier unit 7. This can be... Figure 1 AI system 5 is trained in the image. Image D (image data D) is input into AI system 5, and classification C is generated. This classification C can be a segmentation of image D.

[0058] Figure 3 An AI system 5 with two classifier units 7, 7' according to the present invention is shown. Basic settings and... Figure 2 Similar to the previous system, the difference lies in that the final layer before the output layer 8 of the pre-classifier model 6 is now the input to two classifier units 7 and 7'. Here, each classifier unit 7 and 7' is trained using the individual training data set T. Therefore, the AI ​​system 5 is optimized for two different tasks.

[0059] Figure 4 The following are examples of training pairs: Figure 1 The diagram shows a block diagram of the training method of AI system 5 for classifying medical images D.

[0060] In step I, the provided training data T includes multiple 2D and / or 3D images D and a ground truth G, wherein the multiple 2D and / or 3D images D illustrate organs in the human body, and the ground truth G includes the classification of the organs in image D. As can be seen in the following steps, two different training data sets T1 and T2 are used. In this example, both training data sets T1 and T2 exist within the training data T shown in step I.

[0061] In step II, the pre-classifier model 6 is trained using the first training data set T1 to provide a classification C' of the predefined organ shown in image D based on the baseline ground truth G of the training data T. In this example, the pre-classifier model 6 is designed and trained to generate a feature vector for each pixel of the input image D.

[0062] In step III, classifier unit 7 is trained using a second training data set T2. The training data T2 is input into the trained pre-classifier model 6 so that classifier unit 7 provides a classification C of a predefined organ shown in image D based on the ground truth G of the training data T. In this example, classifier units 7 and 7' are designed to process multiple feature vectors, wherein for each pixel of image D of the training data T, classifier unit 7 is trained by inputting the feature vectors of the pixel and multiple other pixels (particularly neighboring or nearby pixels) at predefined relative positions of the pixel. To achieve this, a predefined region mask M is selected and moved across the training image D. Pixels at predefined positions within region mask M and their ground truth G are then used to train classifier unit 7. The pre-classifier model 6 is no longer trained concurrently with and after classifier unit 7 is trained.

[0063] Figure 5A working example of the invention is shown. In this example, the pre-classifier model 6 is the base classifier model 6. The processing of the AI ​​system 5 is based on a fast classification method based on sparse sampling, which rapidly adapts to new use cases thanks to the fine-tuning of the machine learning strategy by the base classifier model 6 without requiring changes to the base classifier itself.

[0064] Basic classifier model 6 provides organ labels for all organs in the body. Here, the robustness of classification is improved by merging multiple nearby samples and using the most common labels. This is indicated by five arrows from the nearby points of image D to the five boxes indicating the feature vector F. These arrows symbolize the processing of five pixels by basic classifier model 6 before the penultimate layer where the feature vector is taken. These five feature vectors are shown as being connected on the right side of the boxes symbolizing the input to classifier unit 7 (arrows) of basic classifier model 6.

[0065] Therefore, instead of directly using labels, the fundamental features from the final layer of the basic classifier model 6 are processed, allowing new classifiers to be defined based on those features. The classifier units receive far more information than labels alone and can be more easily adapted to desired use cases. The second classifier method outperforms common rule-based labeling methods because it is optimized for task-specific training data.

[0066] In short, the basic classifier model 6 extracts features from a given location and predefined nearby points. These features are then concatenated and fed into the classifier unit, and transformed into the desired output space (classification C).

[0067] Although the present invention has been disclosed in the form of preferred embodiments and variations thereof, it will be understood that numerous additional modifications and variations can be made thereto without departing from the scope of the invention. For clarity, it is to be understood that the use of "a" or "an" throughout this application does not exclude a plurality, and "comprising" does not exclude other steps or elements. The expression "a number" means "at least one". References to "unit" or "device" do not exclude the use of more than one unit or device. Individuals having male or female gender identities are also included in this terminology, independent of the use of grammatical terms.

[0068] Reference tag list 1. Training equipment 2 Data Interface 3 Pre-trained units 4 Classifier Training Unit 5 AI systems 6. Preclassifier Model 7, 7' Classifier Unit 8 Output Layer C, C' Classification D Image data / image F eigenvectors G reference true value M region mask T training data T1 (First) Training Data T2 (second) training data.

Claims

1. A training method for training an AI system (5) for classifying medical images (D), the AI ​​system (5) comprising a machine learning pre-classifier model (6) and machine learning classifier units (7, 7'), wherein several feature vectors (F) of the layer preceding the output layer (8) of the pre-classifier model (6) are the inputs of the classifier units (7, 7'), the training method comprising the following steps: - Provide training data (T), which includes multiple 2D and / or 3D images (D) and a baseline ground truth (G), wherein the multiple 2D and / or 3D images (D) illustrate organs in the human body, and the baseline ground truth (G) includes the classification of the organs in the images (D). - The pre-classifier model (6) is trained using the training data (T) to provide a classification (C') of the predefined organ shown in the image (D) based on the baseline ground truth (G) of the training data (T). - The classifier unit (7, 7') is trained by inputting the training data (T) into the trained pre-classifier model (6) so that the classifier unit (7, 7') provides the classification (C) of the predefined organ shown in the image (D) based on the baseline truth (G) of the training data (T).

2. The training method according to claim 1, wherein two different training data sets (T, T1, T2) are used, wherein the pre-classifier model (6) is trained using the first training data set (T1), and the classifier units (7, 7') are trained using the second training data set (T2), preferably wherein the pre-classifier model (6) is no longer trained while the classifier units (7, 7') are being trained.

3. The training method according to any one of the preceding claims, wherein, in order to train the pre-classifier model (6), the image (D) of the training data (T) is prepared using a sparse sampling descriptor.

4. The training method according to any one of the preceding claims, wherein the pre-classifier model (6) is designed and trained to generate a feature vector (F) for each pixel of the input image (D), and the classifier unit (7, 7') is designed to process a plurality of feature vectors (F), wherein for each pixel of the image (D) of the training data (T), the classifier unit (7, 7') is trained by inputting the feature vectors (F) of the pixel and a plurality of other pixels at a predefined relative position of the pixel, so that the classifier unit (7, 7') provides a classification (C) based on the baseline ground truth (G) of the training data (T), wherein the plurality of other pixels at the predefined relative position of the pixel are, in particular, neighboring or nearby pixels.

5. The training method according to claim 4, wherein the pre-classifier model (6) and / or the classifier unit (7, 7') is trained by: selecting a predefined region mask (M), moving the region mask (M) on the training image (D), and selecting pixels at predefined locations inside the region mask (M) and their ground truth values ​​(G) for processing for training.

6. The training method according to any one of the preceding claims, wherein the training data (T), at least the training data (T) used to train the classifier units (7, 7'), includes random samples.

7. The training method according to any one of the preceding claims, wherein the training data (T), particularly the second training data set (T2) for training the classifier units (7, 7'), includes predefined task-specific localized and labeled images (D).

8. The training method according to any one of the preceding claims, wherein the AI ​​system (5) includes a plurality of machine learning classifier units (7, 7'), wherein the layer preceding the output layer (8) of the pre-classifier model (6) is the input of each classifier unit (7, 7'), preferably wherein each classifier unit (7, 7') is trained using an individual training data set (T).

9. A training apparatus (1) for training an AI system (5) for classifying medical images (D), the AI ​​system (5) comprising a machine learning pre-classifier model (6) and machine learning classifier units (7, 7'), wherein several feature vectors (F) of the layer preceding the output layer (8) of the pre-classifier model (6) are the inputs of the classifier units (7, 7'), the training apparatus (1) comprising: - A data interface (2) is designed to receive training data (T), which includes multiple 2D and / or 3D images (D) and ground truth values ​​(G), wherein the multiple 2D and / or 3D images (D) illustrate organs in the human body, and the ground truth values ​​(G) include the classification of the organs in the images (D). - A pre-training unit (3) is designed to train the pre-classifier model (6) using the training data (T) to provide a classification (C') of the predefined organs shown in the image (D). - Classifier training unit (4), which is designed to train the classifier unit (7, 7') using the training data (T) input into the trained pre-classifier model (6) to provide a classification (C) of the predefined organ shown in the image (D). The training device (1) is designed to train the pre-classifier model (6) and the classifier units (7, 7') using the training method according to any one of the preceding claims.

10. An AI system (5) for classifying medical images (D), the AI ​​system (5) comprising a machine learning preclassifier model (6) and machine learning classifier units (7, 7'), wherein the layer preceding the output layer (8) of the preclassifier model (6) is the input of the classifier units (7, 7'), wherein the AI ​​system (5) has been trained by a training method according to any one of claims 1 to 8.

11. The AI ​​system (5) according to claim 10, wherein the last or penultimate feature layer before the output layer (8) of the pre-classifier model (6) is the input of the classifier unit (7, 7').

12. The AI ​​system (5) according to claim 10 or 11, wherein the classifier unit (7, 7') is smaller than the pre-classifier model (6), preferably wherein the AI ​​system (5) includes a plurality of machine learning classifier units (7, 7'), wherein the layer before the output layer (8) of the pre-classifier model (6) is the input of each classifier unit (7, 7').

13. A medical system, particularly a medical imaging system or diagnostic system, comprising the AI ​​system (5) according to any one of claims 10 to 12.

14. A computer program product comprising instructions that, when executed by a computer, cause the computer to perform the training method according to any one of claims 1 to 8.

15. A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the training method according to any one of claims 1 to 8.

Images